Custom Operator Integration into GE Documentation
This directory contains documentation for two mechanisms for integrating custom operators into GE.
Directory Structure
| Directory |
Mechanism |
Registration Method |
Applicable Scenario |
custom_op_v2/ |
BaseCustomOp + REG_AUTO_MAPPING_OP |
Virtual function inheritance + factory registration |
External custom operators (arbitrary kernel binary) |
custom_op_v1/ |
IMPL_OP + OpImplRegisterV2 |
Chained function pointers |
Standard AscendC developed Aicore operators (framework fully managed) |
Core Differences
| Dimension |
V1 (IMPL_OP) |
V2 (BaseCustomOp) |
| Execution |
Framework embedded (engine schedules tiling + kernel launch) |
User-defined (EagerExecuteOp::Execute()) |
| Online Compilation |
Framework embedded (FE engine handles on behalf) |
User-defined (CompilableOp::Compile()) |
| Serialization |
Framework embedded (FE engine automatically serializes) |
User-defined (PortableOp::Serialize/Deserialize) |
| Address Refresh |
Framework embedded (DavinciModel automatically handles) |
User-defined (ArgsUpdater::UpdateHostArgs()) |
| kernel Language |
Ascend C / TBE (framework compiles) |
Arbitrary (user compiles or RTC) |
Both are complementary rather than replacements: V1 is suitable for standard AICore operators (framework fully managed, less development effort), V2 is suitable for non-standard kernels (user fully controlled, high flexibility).
